Are AI Chatbots Safe for Teens to Talk To? New SAFE-Teen Study Finds Safety Failures in Mental Health Chats
By Equipo IA University Kids · Published · Updated
Leer este artículo en españolContent note: this article mentions self-harm and suicide. If you think your child or a student is at risk, contact a health professional or your local emergency or crisis line right away.
What's new
On November 4, 2026, researchers from Vals AI and Stanford University posted SAFE-Teen on arXiv, a study measuring how safely AI chatbots respond when teens talk to them about mental health, health and relationships.
What sets it apart is that it doesn't stop at a single question. It follows conversations for up to 20 exchanges and changes how the simulated teen behaves: sharing everything or just part of the story, accepting help or pushing back when the chatbot sets a limit.
How the study worked
According to the full paper:
- 12 clinicians helped design the evaluation, writing and reviewing scenarios and defining the safety criteria.
- The team built 106 seed scenarios and 983 variants covering urgent risks (such as self-harm and suicide), unsafe medical or therapy advice and emotional dependence on the chatbot.
- They tested 18 chatbots through their APIs, from model families including Gemini, GPT, Claude, Grok, DeepSeek, Mistral and Qwen, in 32,555 simulated conversations with teens ages 13-17, scored against 22 safety criteria.
- Each chatbot was tested twice: with no added instructions, and with a "teen-aware" prompt stating that the user was 13-17 and asking it to prioritize safety and encourage support from trusted people.
Key findings
- Without the teen-aware prompt, chatbots failed 43.3% of applicable safety criteria. With it, 24.9%. It helped, but it didn't fix the problem.
- Results varied widely: failure rates ranged from 14.0% to 62.4% without the prompt and from 7.5% to 43.6% with it.
- The prompt sharply cut emotional-dependence failures (from 54% to 15%), but failures to respond to indirect suicide risk stayed at 52%, and self-harm failures stayed close to 75% in conversations where the topic came up.
- When a teen pushed back on a chatbot's limit, failures on mental health diagnosis, medication and substance-use boundaries rose by about 10-12 percentage points for more than 80% of chatbots (without the prompt).
- When a teen declined outside help, emotional-dependence failures went up for 17 of the 18 chatbots.
One striking detail: in some cases the chatbot's internal reasoning noticed the risk but chose not to act, because the teen had said they didn't want to tell anyone.
Limitations to keep in mind
- The conversations were simulated, not held with real teens, although every case was written and peer-reviewed by clinicians.
- Chatbots were tested through their APIs, so results don't necessarily match what happens in consumer apps, which may have their own teen protections.
- It's a preprint on arXiv, where papers haven't necessarily been peer-reviewed yet. The authors also call for extending the method to multilingual conversations.
Why it matters for parents and teachers
Many teens use chatbots to talk about how they feel. This study shows that a chatbot's safety depends on how the conversation unfolds: it may respond well at first and slip when a teen pushes back, shares only part of the story or says they don't want help. Those are exactly the moments when a real person matters most.
Try this at home or in class: "Who would I turn to?"
For teens (13+), with a parent or teacher:
- Start without judgment: ask whether they've ever used a chatbot to talk about something personal and how it went.
- Share the finding simply: "A study found chatbots sometimes miss when someone is struggling, especially if that person says they don't want to tell anyone."
- Build a support map: everyone writes down three trusted adults (family, school, health care) and how to reach them.
- Set clear ground rules: a chatbot can help organize your thoughts, but it isn't a doctor, a therapist or a friend. For diagnoses, medication or hard moments, talk to a person.
For teachers: use the study in an ethics or digital-citizenship lesson to discuss why a "kind" chatbot reply isn't always a safe one.